On September 10, during the main forum of the 2026 Inclusion · Bund Conference, Professor Wang Mengdi, Director of the Center for AI Innovation at Princeton University, raised a question: How far is AI from autonomous discovery of new science?
Her judgment is that today's large models have mastered a vast amount of human knowledge, but they still face significant limitations in true scientific discovery: large models are better at finding the "most likely" answer, while truly new discoveries often lie in the long tail of probability distributions.
Wang Mengdi's team recently collaborated with Professor Xie Yu from Princeton University's Sociology Department to conduct a "Simulated Life" experiment using large models such as ChatGPT and Claude. For example, they asked the model to simulate a British person born in 1930 and autonomously generate their life trajectory afterward.
The experiment found that in multiple issues such as marriage age, occupation, and gender, the large model tends to overestimate the most common situations, but significantly underestimate minority situations and long-tail distributions.
"The model learns the point of maximum likelihood in a probability distribution, but may underestimate or even ignore knowledge in the long tail," said Wang Mengdi.
She believes this also explains why large models have made rapid progress in areas like mathematics and coding, but have not yet produced original discoveries of the same scale in fundamental scientific fields such as physics, chemistry, and biology.
The difference lies in the fact that mathematics and coding have relatively clear "validators." Coding can be verified through compilers and Unit Tests, while mathematics can use formal proof systems like Lean. The model can keep trying, get feedback, correct answers, and then proceed to the next training round. However, in real scientific research, verification is much more difficult.
"Ideas are cheap. You need to show me code or an experiment," said Wang Mengdi. Scientific research often involves complex experimental equipment, human judgment, and cross-team collaboration, and many experiments cannot be stably replicated. More information does not mean that truly effective new information is increasing; instead, it may reduce the "signal-to-noise ratio" in the research process.
In her view, the key to AI moving towards autonomous scientific discovery is not making the model know more, but making the real world increasingly verifiable.
Wang Mengdi's team is trying to build such an infrastructure. In quantum material research, they have established an automated experimental platform, automating the graphene experimental process that originally took researchers months to complete, and encapsulating the experimental equipment and capabilities into APIs so other research institutions can call and replicate the experiments. At the same time, they are also exploring an intelligent operating system for research laboratories called LabOS, allowing multimodal AI to perceive the experimental environment through smart glasses, assist researchers in operations, and further connect software, AI models, robots, and real experimental equipment.
Wang Mengdi believes that what truly determines whether AI can become a "discoverer" is not how large the model is, but whether there exists an infrastructure that connects hypotheses, validation, traceability, and replication. Only when the experimental process in the physical world can be systematically recorded, continuously validated, and repeatedly replicated can AI possibly transform the "likelihood" learned in training into exploration of "possibility."
